From 7c60e5339c8c491543a29f4960c423495574bb40 Mon Sep 17 00:00:00 2001
From: Skyler Forge <strictlyskyler@gmail.com>
Date: Sat, 18 Apr 2026 23:10:57 -0700
Subject: [PATCH] fix(agent): also send
 chat_template_kwargs.enable_thinking=false for llama.cpp/vLLM

The custom-provider thinking opt-out at run_agent.py only sets
extra_body.think=false, which is Ollama's field. llama.cpp and vLLM
ignore `think` and look for `chat_template_kwargs.enable_thinking`.

As a result, setting `agent.reasoning_effort: none` on a custom provider
pointed at llama.cpp or vLLM has no effect: thinking-capable models
(Qwen3, etc.) still emit <think> blocks, and any subsequent assistant-
prefill turn (e.g. Hermes' thinking-only continuation path) hits
llama.cpp's prefill-incompatibility 400.

Send both fields. Backends ignore unknown body fields by convention,
so this is additive for Ollama.

Tests added alongside the existing custom-provider think tests.
---
 run_agent.py                      | 17 +++++++++-----
 tests/run_agent/test_run_agent.py | 38 +++++++++++++++++++++++++++++++
 2 files changed, 49 insertions(+), 6 deletions(-)

diff --git a/run_agent.py b/run_agent.py
index 010648809823..64276bba417b 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -7134,17 +7134,22 @@ def _build_api_kwargs(self, api_messages: list) -> dict:
             options["num_ctx"] = self._ollama_num_ctx
             extra_body["options"] = options
 
-        # Ollama / custom provider: pass think=false when reasoning is disabled.
-        # Ollama does not recognise the OpenRouter-style `reasoning` extra_body
-        # field, so we use its native `think` parameter instead.
-        # This prevents thinking-capable models (Qwen3, etc.) from generating
-        # <think> blocks and producing empty-response errors when the user has
-        # set reasoning_effort: none.
+        # Custom provider: opt out of thinking when reasoning is disabled.
+        # Different OpenAI-compat backends use different field names and
+        # ignore unknown fields, so send all the common ones:
+        #   * Ollama:                   think: false
+        #   * llama.cpp / vLLM (Qwen):  chat_template_kwargs.enable_thinking: false
+        # Without this, thinking-capable models (Qwen3, etc.) emit <think>
+        # blocks and trigger llama.cpp's prefill-incompatibility 400 on the
+        # next turn (see llama.cpp tools/server/server-common.cpp).
         if self.provider == "custom" and self.reasoning_config and isinstance(self.reasoning_config, dict):
             _effort = (self.reasoning_config.get("effort") or "").strip().lower()
             _enabled = self.reasoning_config.get("enabled", True)
             if _effort == "none" or _enabled is False:
                 extra_body["think"] = False
+                _ctk = extra_body.setdefault("chat_template_kwargs", {})
+                if isinstance(_ctk, dict):
+                    _ctk["enable_thinking"] = False
 
         if self._is_qwen_portal():
             extra_body["vl_high_resolution_images"] = True
diff --git a/tests/run_agent/test_run_agent.py b/tests/run_agent/test_run_agent.py
index d30445cf459b..3bae47578ff1 100644
--- a/tests/run_agent/test_run_agent.py
+++ b/tests/run_agent/test_run_agent.py
@@ -1088,6 +1088,44 @@ def test_non_custom_provider_unaffected(self, agent):
         kwargs = agent._build_api_kwargs(messages)
         assert kwargs.get("extra_body", {}).get("think") is None
 
+    def test_custom_chat_template_kwargs_on_effort_none(self, agent):
+        """Custom provider with effort=none should also inject
+        chat_template_kwargs.enable_thinking=false (llama.cpp / vLLM)."""
+        agent.provider = "custom"
+        agent.base_url = "http://localhost:11434/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.reasoning_config = {"effort": "none"}
+        messages = [{"role": "user", "content": "hi"}]
+        kwargs = agent._build_api_kwargs(messages)
+        extra = kwargs.get("extra_body", {})
+        assert extra.get("think") is False
+        assert extra.get("chat_template_kwargs", {}).get("enable_thinking") is False
+
+    def test_custom_chat_template_kwargs_on_enabled_false(self, agent):
+        """Custom provider with enabled=false should also inject
+        chat_template_kwargs.enable_thinking=false."""
+        agent.provider = "custom"
+        agent.base_url = "http://localhost:11434/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.reasoning_config = {"enabled": False}
+        messages = [{"role": "user", "content": "hi"}]
+        kwargs = agent._build_api_kwargs(messages)
+        extra = kwargs.get("extra_body", {})
+        assert extra.get("think") is False
+        assert extra.get("chat_template_kwargs", {}).get("enable_thinking") is False
+
+    def test_custom_no_chat_template_kwargs_when_reasoning_enabled(self, agent):
+        """Custom provider with reasoning enabled should NOT inject
+        chat_template_kwargs.enable_thinking=false."""
+        agent.provider = "custom"
+        agent.base_url = "http://localhost:11434/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.reasoning_config = {"enabled": True, "effort": "medium"}
+        messages = [{"role": "user", "content": "hi"}]
+        kwargs = agent._build_api_kwargs(messages)
+        extra = kwargs.get("extra_body", {})
+        assert extra.get("chat_template_kwargs") is None
+
 
 
 class TestBuildAssistantMessage: